Exactly. Short and to the point.
This is my take: For usecases where mistakes don't matter a lot, this is actually useful.
Some examples: generated fiction (mistakes are not mistakes but just an unexpected variation of a story), story-building for a game, and also everywhere you can fix the mistakes yourself (by fact-checking, getting started yourself, or escalating).
Of course this is dangerous if you are not able to catch the mistakes yourself.
I have the feeling: if consumers are powerless then AI is absolutely detrimental. They have to suck up the mistakes and even worse, they might be clueless being fed bullshit.
Some people indistinctly sense that and are afraid.
Isn't this also a way to describe what the press gives us every day? I think the dangers you describe existed long before GPT.
Hah!
Plausible Fiction is fabulous framing. It explains succinctly what annoys me about much prose on the internet, plausible fiction. And it should surprise no one: ChatGPT is trained by the internet.
s/save/safe/
maybe they are testing us
1. someone could make a shitload of money building a tool for learning how to prompt engineer LLMs in a formal course setting. Would be better than most of the GPT grifts currently circulating.
2. the ETHZ writer, while having a generally good grasp, seems to not understand the concept of Attention here. Or that you can do additional training on these models (even just RLHF). I don't think I understand RLHF, Attention, or extra training (new NN layers right?) enough to comment further.
The press I know has a lot of pictures. And the comparison is not unfair at all. Both are Parrots repeating and recombining information from somewhere else, or just inventing new noises if there is nothing to copy.